JP7131440B2 - Arithmetic unit for automobile - Google Patents

Description

The technology disclosed herein belongs to the technical field related to computing devices for automobiles.

Recently, technology for recognizing the environment inside and outside the vehicle using deep learning using a neural network has come to be used for automobiles as well.

For example, Patent Document 1 discloses an estimating device for estimating the state of an occupant with respect to equipment of a vehicle. , an estimating device that estimates the state of the occupant using a model and outputs first information indicating the skeleton position of a specific part of the occupant and second information indicating the state of the occupant with respect to the equipment. is disclosed.

JP 2018-132996 A

By the way, these days, the development of an automatic driving system is promoted nationally. In an automatic driving system, generally, information on the environment outside the vehicle is acquired by a camera or the like, and a route along which the vehicle should travel is calculated based on the information on the environment outside the vehicle that has been acquired. In calculating this route, it is important to recognize the environment outside the vehicle, and the use of deep learning is being studied in this recognition of the environment outside the vehicle.

Here, as a concept of software design of functional safety in the computing device responsible for automatic driving, regarding the application of functional safety of multiple modules that make up the program, for example, in the functional safety standard for automobiles (ISO 26262), hazards are evaluated ASIL (Automotive Safety Integrity Level) is proposed as an index. ASIL requires development at a design level corresponding to each functional safety level from ASIL-A to ASIL-D.

On the other hand, recognition of the environment outside the vehicle and calculation of routes using deep learning are still in a developing state, and are said to remain at the level of ASIL-B. For this reason, in order to satisfy the functional safety level of ASIL-D in automobiles having automatic driving functions, it is not sufficient to configure the arithmetic unit only with functions that utilize deep learning.

The technology disclosed herein has been made in view of the above points, and its purpose is to improve the functional safety level in an arithmetic device having a function that utilizes deep learning.

In order to solve the above problems, the technology disclosed herein performs deep learning based on the output from information acquisition means for acquiring information on the environment outside the vehicle, targeting an automotive arithmetic unit mounted in the vehicle. a route calculation unit that calculates route candidates based on the estimation result of the vehicle environment estimation unit; and a route candidate based on the output from the information acquisition unit. An object identification unit that identifies objects outside the vehicle according to a predetermined rule without using learning, a safety area calculation unit that calculates a safety area based on the result of the object identification unit, and the route calculation. a desired motion determining unit for determining a desired motion of the vehicle upon receiving an output from the unit and an output from the safe region calculating unit, wherein the desired motion determining unit determines whether the route candidate is within the safe region. Sometimes, the route candidate is selected as the route on which the vehicle should travel, and a desired motion of the vehicle is determined such that the vehicle passes through the route candidate while the route candidate deviates from the safe area. In some cases, the route candidate is not selected as the route on which the vehicle should travel.

According to this configuration, apart from estimating the environment outside the vehicle by deep learning, objects outside the vehicle are identified according to a predetermined rule. Here, the predetermined rule is a method for certifying target objects, etc., which has been conventionally adopted for automobiles, etc., and the target certification function based on the predetermined rule is at a functional safety level equivalent to ASIL-D. be. Therefore, it can be said that the safe area calculated based on the result of the object identifying unit is an area with high safety.

Then, the desired motion determination unit determines that when a route candidate calculated based on the environment outside the vehicle estimated by deep learning (hereinafter referred to as an estimated route candidate) is within the safe area, the vehicle should travel along the estimated route candidate. Let it be the route. This allows the automobile to travel on a highly safe route. On the other hand, when the estimated route candidate deviates from the safe area, the route candidate is not selected as the route on which the vehicle should travel. At this time, for example, the route passing through the safe area is recalculated based on the object information certified by a predetermined rule, or if the driver is in the car, the driver is left to drive the car. This allows the vehicle to travel on a highly safe route.

Therefore, the functional safety level can be improved in an arithmetic device having a function using deep learning.

In one embodiment of the computing device for a vehicle, the route calculation unit calculates a plurality of route candidates, and the desired motion determination unit calculates a route that deviates from the safe area among the plurality of route candidates, It is not selected as the route on which the vehicle should travel.

According to this configuration, the vehicle can more effectively travel on a highly safe route. For example, when some candidates out of a plurality of route candidates deviate from the safe region and the remaining candidates are within the safe region, one of the remaining routes is selected as the route on which the automobile should travel. The selection allows the car to travel on a safer route. Further, when all of the plurality of route candidates deviate from the safe area, as described above, the route passing through the safe area is recalculated based on the object information certified by the predetermined rule, or the vehicle is driven. When a person is on board, the vehicle can travel on a highly safe route by, for example, leaving the driving to the driver.

In the arithmetic device for automobile, the route calculation unit may be configured to calculate the route candidate using reinforcement learning.

According to this configuration, for example, if safety (for example, low risk of collision with obstacles) is set as the Q value of reinforcement learning, when deep learning is appropriate, an estimated route candidate with high safety is easier to calculate. This allows the vehicle to travel more effectively on a safer route.

As described above, according to the technology disclosed herein, it is possible to improve the functional safety level in an arithmetic device having a function using deep learning.

1 is a block diagram showing a functional configuration of an automotive computing device according to Embodiment 1; FIG. It is a figure which shows an example of the driving|running route set by a calculating|arithmetic apparatus. It is a figure which compares the safe area set by the 1st calculating part, and the safe area set by the 2nd calculating part. FIG. 4 is a diagram showing an example of the relationship between route candidates calculated based on the environment outside the vehicle estimated by deep learning and the safe area. FIG. 9 is another diagram showing the relationship between route candidates calculated based on the environment outside the vehicle estimated by deep learning and the safe area. 4 is a flow chart for determining the driving route of an automobile. FIG. 7 is a block diagram showing the functional configuration of an automotive computing device according to Embodiment 2;

Exemplary embodiments are described in detail below with reference to the drawings.

(Embodiment 1)
FIG. 1 shows the configuration of an automotive computing device (hereinafter simply referred to as computing device 100) according to the first embodiment. The computing device 100 is, for example, a computing device mounted on a four-wheeled automobile 1 . The automobile 1 is an automobile capable of manual driving in which the driver operates the accelerator or the like, assist driving in which the operation is assisted by the driver, and automatic driving in which the driver does not operate the vehicle. be. In the following description, the automobile 1 on which the arithmetic device 100 is mounted is sometimes referred to as the own vehicle 1 in order to distinguish it from other vehicles.

The arithmetic device 100 is a microprocessor composed of one or a plurality of chips, and has a CPU, memory, and the like. Note that FIG. 1 shows a configuration for exerting a function (route generation function described later) according to the present embodiment, and does not show all the functions of the arithmetic device 100 .

As shown in FIG. 1, the arithmetic unit 100 determines a target motion of the automobile 1 based on outputs from a plurality of sensors and the like, and controls the operation of the devices. The sensors and the like that output information to the arithmetic unit 100 include a plurality of cameras 50 that are provided on the body of the automobile 1 or the like and that capture the environment outside the vehicle, and a plurality of cameras that are provided on the body or the like of the automobile 1 and detect targets outside the vehicle. a radar 51, a vehicle speed sensor 52 for detecting the absolute speed of the automobile 1, an accelerator opening sensor 53 for detecting the depression amount of the accelerator pedal of the automobile 1, and a rotation angle (steering angle) of the steering wheel of the automobile 1. The position of the automobile 1 (vehicle position information) is detected using a steering angle sensor 54, a brake sensor 55 that detects the amount of depression of the brake pedal of the automobile 1, and a global positioning system (GPS). and a position sensor 56 for On the other hand, objects controlled by the arithmetic device 100 include the engine 10 , the brake 20 , and the steering 30 .

Each camera 50 is arranged so as to photograph the surroundings of the automobile 1 horizontally at 360 degrees. Each camera 50 captures an optical image representing the environment outside the vehicle and generates image data. Each camera 50 outputs the generated image data to the arithmetic device 100 . The camera 50 is an example of information acquisition means for acquiring information on the environment outside the vehicle.

As with the camera 50, each radar 51 is arranged so that the detection range extends horizontally around the automobile 1 by 360 degrees. The type of radar 51 is not particularly limited, and for example, millimeter wave radar or infrared radar can be adopted. The radar 51 is an example of information acquisition means for acquiring information on the environment outside the vehicle.

The engine 10 is a power drive source and includes an internal combustion engine (gasoline engine, diesel engine). The computing device 100 outputs an engine output change signal to the engine 10 when the automobile 1 needs to be accelerated or decelerated. The engine 10 is controlled by the amount of operation of the accelerator pedal by the driver during manual operation, but is controlled based on the target motion calculated by the arithmetic unit 100 during assist operation or automatic operation. Although not shown, the rotating shaft of the engine 10 is connected to a generator that generates power using the output of the engine 10 .

Brake 20 is an electric brake here. The arithmetic device 100 outputs a brake request signal to the brake 20 when the vehicle 1 needs to be decelerated. Upon receiving the brake request signal, the brake 20 operates a brake actuator (not shown) based on the brake request signal to decelerate the vehicle 1 . The brake 20 is controlled by the amount of operation of the brake pedal by the driver during manual driving, but is controlled based on the target motion calculated by the arithmetic unit 100 during assist driving or automatic driving.

The steering 30 is an EPS (Electric Power Steering) here. Arithmetic device 100 outputs a steering direction change signal to steering 30 when it is necessary to change the traveling direction of automobile 1 . The steering 30 is controlled by the operation amount of the steering wheel (so-called steering wheel) of the driver during manual driving, but is controlled based on the target motion calculated by the arithmetic unit 100 during assist driving or automatic driving.

The transmission 40 is a stepped transmission. Arithmetic device 100 outputs a gear change signal to transmission 40 in accordance with the driving force to be output. During manual driving, the transmission 40 is controlled by the driver's operation of the shift lever, the driver's operation amount of the accelerator pedal, and the like. controlled.

Arithmetic device 100 transmits a control signal based on the output of accelerator opening sensor 52 or the like to engine 10 or the like during manual operation. On the other hand, the arithmetic unit 100 sets the travel route of the automobile 1 during assisted driving or automatic driving, and outputs a control signal to the engine 10 or the like so that the automobile 1 travels along the travel route. The arithmetic unit 100 includes a first arithmetic unit 110 having a function of estimating the environment outside the vehicle using deep learning as a configuration for generating a control signal to be output to the engine 10 or the like during assisted driving or automatic driving. , a second calculation unit 120 having a function of recognizing an object based on a predetermined rule without using deep learning, and receiving outputs from the first and second calculation units 110 and 120 to calculate a desired motion of the automobile 1. A target motion determination unit 130 for determination, and an energy management unit that calculates control amounts for various devices (such as the engine 10 described later) so as to maximize energy efficiency in achieving the target motion determined by the target motion determination unit 130. 140.

Based on the output from the camera 50 and the radar 51, the first calculation unit 110 includes an external environment estimation unit 111 that estimates the vehicle external environment using deep learning, and the external environment estimated by the external environment estimation unit 111. On the other hand, the first safety area setting unit 112 that sets the first safety area SA1 (see FIG. 3), the estimation result of the vehicle external environment estimation unit 111, and the first safety area SA1 set by the first safety area setting unit 112 and a first route calculation unit 113 for calculating a first route candidate based on.

The vehicle exterior environment estimation unit 111 estimates the vehicle exterior environment by image recognition processing using deep learning based on the outputs from the camera 50 and the radar 51 . Specifically, the vehicle exterior environment estimation unit 111 constructs object identification information by deep learning based on the image data from the camera 50, and integrates positioning information by radar into the object identification information to create a 3D image representing the vehicle exterior environment. Create maps. In addition, an environment model is generated by integrating behavior prediction of each object based on deep learning with respect to the 3D map. In deep learning, a multilayer neural network (DNN: Deep Neural Network) is used. As a multilayer neural network, for example, there is a CNN (Convolutional Neural Network).

First safety area setting unit 112 sets first safety area SA1 for the 3D map created by vehicle exterior environment estimation unit 111 . This first safety area SA1 is set as an area through which the own vehicle can pass using a model constructed using deep learning. The model is constructed, for example, by reconstructing a model constructed in advance for each vehicle type of the automobile 1 based on the driver's past driving history and the like. The first safety area SA1 is basically on the road, and there are dynamic obstacles such as other vehicles and pedestrians, and static obstacles such as the center separator and the center pole. It refers to the area that does not The first safety area SA1 may include a shoulder space for emergency parking.

The first route calculation unit 113 calculates a first route candidate that passes through the first safety area SA1 set by the first safety area setting unit 112 . The first route calculation unit 113 calculates first route candidates using reinforcement learning. In reinforcement learning, an evaluation function is set for the results of a series of simulations (here, route candidates), and a good evaluation is given to simulations that meet a certain purpose, and a low evaluation is given to results that do not. , is a function to learn route candidates that meet the purpose. An actual calculation method will be described later.

The second computing unit 120 includes an object recognizing unit 121 that recognizes an object outside the vehicle according to a predetermined rule without using deep learning based on outputs from the camera 50 and the radar 51, and an object recognizing unit 121. 121, a second safety area calculation unit 122 that calculates a second safety area SA2 (see FIG. 4), the result of the object identification unit 121, and the second safety area set by the second safety area setting unit 122. and a second route calculator 123 that calculates a second route candidate based on the safe area SA2.

The object certifying unit 121 certifies an object based on a conventionally used object certifying rule. The object is, for example, a traveling vehicle, a parked vehicle, a pedestrian, or the like. The object recognition unit 121 also recognizes the relative distance and relative speed between the own vehicle and the object. Moreover, the object recognition unit 121 also recognizes the traveling road (including lane markings and the like) based on the outputs from the camera 50 and the radar 51 .

The second safe area setting unit 122 sets the second safe area SA2 as an area in which collision with the object recognized by the object recognition unit 121 can be avoided. This second safe area SA2 is set based on a predetermined rule such as, for example, regarding the surrounding number m of an object as an unavoidable range. The second safety area setting unit 122 is configured to be able to set the second safety area SA2 in consideration of the speed of the traveling vehicle and the speed of pedestrians. Like the first safety area SA1, the second safety area SA2 is basically on the road and is free from dynamic obstacles such as other vehicles and pedestrians, as well as the central separator and the center pole. An area where there are no static obstacles such as The second safety area SA2 may include a shoulder space for emergency parking.

The second route calculation unit 123 calculates second route candidates that pass through the second safety area SA2 set by the second safety area setting unit 122 . An actual calculation method will be described later.

The target motion determination unit 130 receives outputs from the first and second calculation units 110 and 120, particularly information on the first and second safety areas SA1 and SA2 and information on the first and second route candidates. to determine the target motion of the automobile 1. The target motion determination unit 130 sets a route for the automobile 1 to travel, and issues requests to various devices (mainly the engine 10, the brakes 20, the steering 30, and the transmission 40) so that the automobile 1 travels along the route. An actuation amount (for example, an engine torque, an actuation amount of a brake actuator, etc.) is determined.

The energy management unit 140 calculates control amounts for various devices (such as the engine 10 to be described later) so as to achieve the highest energy efficiency in achieving the required amounts for the various devices determined by the target exercise determination unit 130.

For example, when outputting the required driving force determined by the desired motion determination unit 130, the energy management unit 140 adjusts the number of gears of the transmission 40 and intake/exhaust valves (not shown) so that the fuel consumption of the engine 10 is minimized. ) and fuel injection timing of an injector (not shown). In addition, when outputting the target braking force, the energy management unit 140 increases the regenerative power generation amount of the generator connected to the engine 10 and the driving load of the cooling compressor so that the engine brake is minimized. generating the braking force; Also, the energy management unit 140 controls the vehicle speed and the steering angle so that the rolling resistance applied to the automobile 1 during cornering is minimized. Specifically, the generation of the braking force and the steering timing are adjusted so that the rolling and the pitching in which the front portion of the automobile 1 sinks are synchronously generated to produce a diagonal roll posture. Due to the diagonal roll posture, the load applied to the front turning outer wheel increases, enabling turning with a small steering angle and reducing the rolling resistance applied to the automobile 1 .

Next, a method for calculating the first route candidate will be described with reference to FIG. Although illustration of the first safety area SA1 is omitted in FIG. 2, the calculated first route candidate is a route passing through the first safety region SA1. The calculation of the first route candidate is performed when the driving mode of the automobile 1 is assist driving or automatic driving, and the calculation of the first route candidate is not performed when the driving mode is manual driving.

First, as shown in FIG. 2, the first route calculation unit 121 executes grid point setting processing based on the travel path information. In the grid point setting process, the ECU 10 first specifies the shape of the road 5 (that is, the direction in which the road 5 extends, the width of the road, etc.), and places grid points Gn (n=1, 2, . . . ) on the road 5. . . N) is set.

The grid area RW extends from the periphery of the vehicle 1 to a predetermined distance ahead of the vehicle 1 along the travel path 5 . This distance (longitudinal length) L is calculated based on the current vehicle speed of the vehicle 1 . In this embodiment, the distance L is the expected distance traveled at the current vehicle speed (V) in a predetermined fixed time t (eg, 3 seconds) (L=V×t). However, the distance L may be a predetermined fixed distance (eg 100m) or it may be a function of vehicle speed (and acceleration). Also, the width W of the grid area RW is set to the width of the running path 5 .

The grid area RW is divided into a large number of rectangular grid sections by a plurality of grid lines extending along the extending direction X and the width direction (horizontal direction) Y of the travel path 5 . Grid points Gn are intersections of grid lines in the X direction and the Y direction. The X-direction spacing and the Y-direction spacing of the grid points Gn are each set to a fixed value. In this embodiment, for example, the grid spacing in the X direction is 10 m, and the grid spacing in the Y direction is 0.875 m. However, the grid interval may be a variable value according to the vehicle speed or the like.

In FIG. 2, the running path 5 is a straight section, and the grid area RW and the grid sections are set in a rectangular shape. If the road includes curved sections, the grid areas and grid sections may or may not be rectangular.

Next, in the calculation of the first route candidate, the first route calculation unit 113 sets a predetermined grid point GT within the grid area RW as the target arrival position PE ( Set the target speed in GT). The external signal is, for example, a guidance instruction signal to a destination (parking area, etc.) transmitted from a navigation system (not shown) mounted on the vehicle 1 .

Next, the first route calculation unit 113 executes route candidate calculation processing. In this route candidate calculation process, the first route calculation unit 113 first executes a route candidate calculation process for calculating a plurality of first route candidates R1m (m=1, 2, 3, . . . ). This process is similar to the route candidate calculation using the conventional state lattice method.

An outline of route candidate calculation will be described. The first route calculator 113 creates route candidates from the current position PS (start point) of the vehicle 1 to each grid point Gn (end point) within the grid area RW. Also, the first route calculation unit 113 sets speed information at the end point.

The start and end points are connected via one or more grid points Gn or not via grid points Gn. For each first route candidate R1m, the first route calculation unit 113 calculates position information by fitting a route curve pattern between grid points, and calculates a speed change profile to match the speed change pattern. The speed change pattern includes rapid acceleration (eg, 0.3 G), slow acceleration (eg, 0.1 G), vehicle speed maintenance, slow deceleration (eg, -0.1 G), and rapid acceleration (eg, -0.3 G). It is generated in combination and set over a predetermined length (for example, 50 m to 100 m) of the first route candidate R1m, not for each grid.

Furthermore, the first route calculation unit 113 sets sampling points SP for each route candidate and calculates speed information at each sampling point SP. FIG. 2 shows only three first route candidates R11, R12, R13 among many route candidates. Note that, in the present embodiment, each first route candidate R1m is a route from the starting point to an arrival position after a fixed time (for example, 3 seconds).

Next, the first route calculation unit 113 calculates the route cost of the obtained first route candidate R1m. In calculating the route cost of each first route candidate R1m, the first route calculation unit 113 calculates, for example, the inertial force Fi due to the motion of the vehicle 1, the obstacle (here, the other vehicle 3) and Calculate the collision probability Pc and the impact force Fc (or reaction force to the collision) that the occupant receives due to this collision, calculate the external force FC that acts on the occupant based on these values, and A total value of the external forces FC (absolute values) at the sampling points SP can be calculated as the route cost (candidate route cost) EPm of the first route candidate R1m.

Then, the first route calculation unit 113 outputs all the first route candidates R1m together with their respective route cost information to the desired motion determination unit 130 .

Route candidates are set as described above. The second route candidate is also calculated basically in the same manner as described above. However, for the second route candidate, the route with the lowest route cost is selected and output to the desired motion determination unit 130 .

Here, in the first embodiment, the route candidates include the first route candidate calculated by the first calculation unit 110 and the second route candidate calculated by the second calculation unit 120. However, the desired motion determination unit 130 basically adopts the first route candidate. This is because the first route candidate set using deep learning or reinforcement learning is a route that reflects the driver's intentions, that is, it is difficult for the driver to avoid redundancy such as being too cautious in avoiding obstacles. This is because it tends to be a path that does not make you feel.

However, recognition of the environment outside the vehicle using deep learning is still in a developing state. In other words, an environmental model built using deep learning can calculate accurate information in a range similar to the information on which the environmental model is based. , the environment outside the vehicle may be estimated to deviate from the actual environment outside the vehicle.

For example, FIG. 3 shows a first safety area SA1 set by the first safety area setting unit 112 and a second safety area SA2 set by the second safety area setting unit 122. As shown in FIG. The first safety area SA1 is the hatched portion in FIG. 3, and the second safety area SA2 is the hatched portion in FIG. 3 excluding the inside of the dotted frame. As shown in FIG. 3, in the first safety area SA1, part of the other vehicle 3 is also set in the safety area. This may occur, for example, when the vehicle width of the other vehicle 3 cannot be accurately estimated in image recognition by deep learning.

In this way, certification of the vehicle environment using deep learning may estimate the environment outside the vehicle that deviates from the actual environment outside the vehicle. ASIL), functions using deep learning are considered to be equivalent to ASIL-B. Therefore, it is necessary to devise ways to improve the functional safety level.

Therefore, in the first embodiment, the desired motion determining unit 130 of the arithmetic device 100 determines that the automobile 1 (self-vehicle 1) travels along the first route candidate when the first route candidate is within the second safety area SA2. While determining the desired motion of the vehicle 1 so that the vehicle 1 passes through the first candidate route, when the first candidate route deviates from the second safety area SA2, the first route candidate The candidate is not selected as the route on which the automobile 1 should travel. More specifically, when calculating the first route candidates, the desired motion determination unit 130 selects all of the plurality of first route candidates R1n (n=1, 2, 3, . . . ) set as route candidates. deviates from the second safe area SA2, the second route candidate is selected as the route on which the vehicle 1 should travel.

For example, as shown in FIG. 4, it is assumed that first route candidates R11, R12, and R13 are set by the first route calculation unit 113. FIG. Among the three first route candidates R11 to R13, routes R11 and R12 partially deviate from the second safety area SA2, but route R13 passes through the second safety area SA2. ing. At this time, the desired motion determination unit 130 does not select the routes R11 and R12 that deviate from the second safety area SA2 as the route on which the vehicle 1 should travel, but selects the route R13 as the route on which the vehicle 1 should travel. .

On the other hand, it is assumed that first route candidates R14, R15, and R16 are set by the first route calculation unit 113 as shown in FIG. All of the three first route candidates R14 to R16 deviate from the second safety area SA2. At this time, the desired motion determination unit 130 selects the second route candidate R2 having the lowest route cost among the plurality of second route candidates calculated by the second calculation unit 120 as the route on which the automobile 1 should travel.

Since the object certifying unit 121 certifies the object based on a conventionally existing predetermined rule, it is possible to accurately certify the size of the object and the like. In addition, since the second safe area setting unit 122 sets the second safe area SA2 based on a predetermined rule such as regarding the surrounding number m of the object as the unavoidable range, the second route candidate is the other vehicle 3 Even in the case of avoiding the That is, the function of the second calculation unit 120 can be equivalent to ASIL-D. Therefore, when desired motion determination section 130 calculates the first route candidates, if all of the plurality of first route candidates R1n set as route candidates deviate from second safety area SA2, the second By selecting a route candidate as the route on which the automobile 1 should travel, the automobile 1 can travel a highly safe route. As a result, the functional safety level can be improved in the arithmetic device 100 having the function of using deep learning.

Next, the processing operation of the arithmetic device 100 when determining the driving route of the automobile 1 will be described with reference to the flowchart of FIG. The driving mode of the automobile 1 is assist driving or automatic driving.

First, in step S1, the arithmetic unit 100 reads information from each camera 50, each radar 51, and each sensor 52-56.

Next, the arithmetic unit 100 performs calculation of the first route candidate and calculation of the second route candidate in parallel.

In step S2, the computing device 100 estimates the environment outside the vehicle using deep learning.

In the next step S3, the arithmetic device 100 sets the first safety area SA1.

In the next step S4, the computing device 100 uses reinforcement learning to calculate the first route candidates.

On the other hand, in step S5, the computing device 100 authenticates the object based on a predetermined rule.

In step S6, the arithmetic device 100 sets the second safety area SA2.

At step S7, the arithmetic device 100 calculates the second route candidate based on a predetermined rule.

Subsequently, in step S8, the arithmetic device 100 outputs the information of steps S4 and S5 to the desired motion determination unit 130, and determines whether the first route candidate is within the second safety area SA2 in the desired motion unit 130. determine whether In this step S8, when the plurality of first route candidates includes all routes included in the second safety area SA2 YES, the process proceeds to step S9, while all of the plurality of first route candidates If at least a portion of the route deviates from the second safety area SA2 and NO, the process proceeds to step S10.

In step S9, the desired motion determination unit 130 selects the first route candidate as the route along which the automobile 1 should travel.

On the other hand, in step S10, the desired motion determination unit 130 selects the second route candidate as the route along which the automobile 1 should travel.

In the next step S11, the desired motion determining section 130 calculates the desired motion of the automobile 1. FIG.

In subsequent step S12, the energy management unit 140 sets the target control amount such that the target motion calculated in step S11 is achieved and the energy efficiency is maximized.

In the next step S13, the arithmetic unit 100 controls the operation of each device so that the control amount of each device becomes the target control amount calculated in step S12. After step S13, the process returns.

Therefore, in the first embodiment, the computing device 100 includes the vehicle exterior environment estimation unit 111 that estimates the vehicle exterior environment using deep learning based on the output from the information acquisition means that acquires the vehicle exterior environment information, and the vehicle exterior environment estimation unit 111 that estimates the vehicle exterior environment. Based on the estimation result of the environment estimation unit 111, the first route calculation unit 113 calculates the first route candidate, and based on the output from the information acquisition unit, without using deep learning, according to a predetermined rule Based on the result of the target object recognition unit, the object recognition unit 121 that recognizes the object outside the vehicle, the second safety region calculation unit 122 that calculates the safety region, and the output from the first route calculation unit 113 and the first A desired motion determination unit 130 that receives the output from the 2-safety area calculation unit 122 and determines the desired motion of the vehicle 1 . When the first route candidate is within the second safe area SA2, the desired motion determining unit 130 selects the first route candidate as the route on which the vehicle 1 should travel, and the vehicle 1 passes through the first route candidate SA1. While the desired motion of the vehicle 1 is determined as described above, when the route candidate deviates from the safe area, the route candidate is not selected as the route on which the vehicle 1 should travel.

According to this configuration, it is possible to improve the functional safety level in the arithmetic device 100 having the function of using deep learning.

(Embodiment 2)
Hereinafter, Embodiment 2 will be described in detail with reference to the drawings. In the following description, the same reference numerals are assigned to the same parts as in the first embodiment, and detailed description thereof will be omitted.

In the second embodiment, the configuration of the second arithmetic unit 220 in the arithmetic device 200 is different from that of the arithmetic device 100 in the first embodiment. Specifically, in the second embodiment, unlike the first embodiment, the second calculation unit 220 does not include the second route calculation unit.

Even with this configuration, when all of the first route candidates calculated by the first calculation unit 110 deviate from the second safety area SA2, the desired motion determination unit 130 determines the first route candidates as follows: Do not select it as a route to travel. In the second embodiment, when the desired motion determination unit 130 does not select the first route candidate, for example, the driving mode of the vehicle 1 is switched to manual driving so that the vehicle 1 can travel a safe route. do. At this time, when switching from automatic operation to manual operation, it is desirable to inform the driver of the change in operation mode by means of a buzzer or the like.

Therefore, in the second embodiment as well, the functional safety level can be improved in the arithmetic device 200 having the function of using deep learning.

(Other embodiments)
The technology disclosed herein is not limited to the above embodiments, and substitutions are possible without departing from the scope of the claims.

For example, in Embodiments 1 and 2 described above, after the target motion determination unit 130 determines the target motion, the energy management unit 140 sets the target control amount so that the target motion is achieved and the energy efficiency is maximized. rice field. Alternatively, the energy management unit 140 may be omitted. That is, the desired control amount that achieves the desired exercise may be set by the desired exercise determination unit 130 .

The above-described embodiments are merely examples, and should not be construed as limiting the scope of the present disclosure. The scope of the present disclosure is defined by the claims, and all modifications and changes within the equivalent range of the claims are within the scope of the present disclosure.

The technology disclosed herein is useful as an automotive computing device mounted on an automobile.

1 Automobile 100 Automobile computing device 111 External environment estimation unit 113 First route candidate calculation unit 121 Object identification unit 122 Second safe area setting unit 130 Desired motion determination unit

Claims (3)

An automotive computing device mounted in an automobile,
a vehicle exterior environment estimating unit that estimates the vehicle exterior environment using deep learning based on the output from the information acquisition means that acquires information on the vehicle exterior environment;
a route calculation unit that calculates a route candidate based on the estimation result of the vehicle external environment estimation unit;
an object recognizing unit that certifies an object outside the vehicle according to a predetermined rule without using deep learning, based on the output from the information acquisition means;
a safety area calculation unit that calculates a safety area based on the result of the object identification unit;
a desired motion determination unit that receives an output from the route calculation unit and an output from the safe area calculation unit and determines a desired motion of the vehicle;
The desired motion determination unit selects the route candidate as a route on which the vehicle should travel when the route candidate is within the safe area, and sets the target motion of the vehicle so that the vehicle passes through the route candidate. is determined, and when the route candidate deviates from the safe area, the route candidate is not selected as the route on which the vehicle should travel. The automotive computing device according to claim 1,
The route calculation unit calculates a plurality of route candidates,
A computing device for a vehicle, wherein the desired motion determination unit does not select a route that deviates from the safe region from among the plurality of route candidates as a route on which the vehicle should travel. 3. In the automotive computing device according to claim 1 or 2,
The arithmetic unit for an automobile, wherein the route calculation unit calculates the route candidate using reinforcement learning.

Images